Robots are notoriously lousy at coping with unfamiliar situations. One of the major obstacles to having robots assist us at home is that they’re going to need to be flexible, and home users aren’t going to want to sit down and write a new servo control subroutine every time their robot needs to perform a task for which it has not been preprogrammed. The robot in the video above is able to learn via observation and replication, which is how we humans figure things out. The HOAP-3 robot, made by Fujitsu, has been programmed by Sylvain Calinon and other researchers at the Swiss Federal Institute of Technology in Lausanne to watch how a human performs an action, and then duplicate that action. If the robot doesn’t get it right, it can learn on the fly as a human corrects its motions, which (if you think about it) is pretty much how you were taught to write, throw a baseball, ride a bike, and so forth. I’d much rather have a dumb robot with the ability to learn than a smart robot that I can’t teach new things to, especially if the teaching process is no more complicated than a game of Simon says. Er, make that Evan says.

[ Sylvain Calinon’s Homepage ] VIA [ New Scientist ]

Well, I’m disappointed. It looks like a drone dogfight, which would be the only thing more robotic than the acting in Top Gun, is not going to be feasible for at least the next 25 years, according to the Pentagon… Dyke Weatherington, deputy director for the Pentagon’s Unmanned Aircraft Systems (UAS) Task Force, states that “there’s really no way that a system that’s remotely controlled can effectively operate in an offensive or defensive air combat requirement. The requirement of that is a fully autonomous system, we don’t have that level of autonomy yet and frankly in the roadmap that will take many years to get to.” He’s got a point there, sort of, but I also think that the speed of development of aerial combat drones is being vastly underestimated, especially when it comes to the AI. I’m not saying I think it’ll happen in the next year, but 25 years is a very, very long time when it comes to computers. I mean, 25 years ago was what, 1982? Think about the amount of change that has happened between then and now, and project it forward. We’re rapidly approaching the point where hardware is not going to be the limiting factor, and where computers will be able to operate (and interoperate) fast enough to make faster and arguably better decisions in an active combat environment than humans are capable of. The hardest part, in my opinion, is not going to be developing the autonomy: it’s going to be getting past all of those prickly ethical issues that come up when you have robots deciding whether or not to shoot at people.

BTW, the drone matchup above is the Boeing X-45 versus the MiG UCAV.

VIA [ Danger Room ]

Yeah, I admit it, I listen to NPR. On Sunday’s Weekend Edition, there was a story on artificial intelligence, from the 2007 Singularity Summit held in San Francisco in September. Specifically, the story (and the summit) was about what artificial intelligence is, and where it’s likely to go. I was particularly interested in the two separate types of AI; specialized and generalized. It seems like we’ve gotten pretty good with the specialized AI (robots that can perform one task, like vacuuming), and generalized AI is the next necessary step when it comes to evolving adaptive complexity (in the software sense) in robots. So far, I think the best we’ve been able to do is to create robots with lots of specialized AI that kind of pretends to be generalized AI… But maybe that’s a good thing, because whether we really want robots to be making (for lack of a better word) inexperienced decisions is a tough (and potentially scary) question.

Click “Listen Now” at the link below to hear the story. The picture, by the way, shows one possible destiny: humans as food, from The Matrix.

[ NPR Weekend Edition ]

In this somewhat abusive video, the researchers are trying to get a robot named Justin to hit himself, or hit (ahem) his two balls together, but Justin is having none of it: he knows exactly where his arms are in relation to the rest of his body at all times, and is able to prevent collisions. To figure out where the different parts of his body are, Justin uses feedback from the artificial muscles in his arms, which is analogous to how humans do it. When you touch your nose with your eyes closed (or try to), you’re using kinaesthetics to know where your arm, hand, and finger is without looking. Instead of relying on pre-programming, Justin is able to self-model his appendages to dynamically prevent accidents.

If you’d like to know more about Justin, please see the citation in the comments or visit http://www.phriends.eu/.

VIA [ New Scientist ]

One of the most annoying things about airplanes and helicopters is their dependence on a suitable place to land. Of course, helicopters are much more versatile than airplanes are, but they still need a large, reliably flat and stable surface. Landing any sort of aircraft on a boat, much less a moving truck, is dangerous under the best of circumstances and usually impossible most of the rest of the time. It’s becoming clear how much precision autopilots are capable of, and researchers at Georgia Tech and MIT are taking advantage of robots’ skill at flying to perform acrobatics designed to make drone launch and recovery much more versatile.

This video shows a small, autonomous helicopter developed by Georgia Tech landing on slopes ranging from 45 up to 60 degrees. Although the surface that the helicopter lands on is velcro, it’s possible to have the helicopter reverse the pitch of its rotors on landing, which would push it down onto the landing pad. Notice how the helicopter is able to recover from a missed landing in the middle of the video. Skills like this, combined with adaptive sensors, would allow a drone to land virtually anywhere, even on a non-level moving surface.

Both an airplane propeller and a helicopter rotor operate on the same principle: air moving rapidly over a spinning, curved surface generates thrust. It turns out that if you’ve got a small enough plane and a big enough engine (i.e. a drone), you can actually make a helicopter out of an airplane, using the propeller as a rotor. It’s tricky and requires a lot of fine manipulation of the wing and tail control surfaces, but autopilots excel at fine manipulation, and MIT has been able to create an autopilot program to transition an airplane from vertical takeoff, to horizontal flight, to a vertical landing:

The US Air Force actually messed around with a concept sort of like this back in the 50s. Although it was jet powered, the X-13 also took off and landed from a suspension rack. It’s worth noting that it was nearly impossible for the human pilot to accomplish this maneuver.

VIA [ New Scientist ]

The DARPA Urban Challenge for robotic vehicles concluded on Saturday, with Carnegie Mellon’s robotic 2007 Chevy Tahoe, named Boss, taking top honors and the $2 million first prize. In contrast to the first DARPA Grand Challenge in 2004, where no robots completed the course, there were six finishers this year, although the last few took longer than the six hour time limit. Still, it’s a pretty impressive feat for a robotic car to spend six hours on the road completing tasks among other cars with no human guidance whatsoever. Here’s some short clips from the event; hopefully we’ll see more video in not too long:

The critical point here is that all of the robots were able to make good decisions when confronted with complicated variables and rapidly changing situations. The top three committed NO traffic violations over the 60 mile course. The deciding factor then became speed, and at an average speed of 14 mph, Boss won out by a 20 minute margin over second place Stanford (average speed 13 mph), winners of the first DARPA Grand Challenge in 2005.

[ DARPA Urban Challenge ] VIA [ Danger Room ]

The qualifiers for the DARPA Urban Challenge started on October 26, with 35 teams competing for 20 slots in the finals, which will start November 3rd. If you’re not familiar with the DARPA Grand Challenge series, the first successful one took place in 2005. It was an off road race for driverless vehicles over a 132 mile stretch of rugged Mojave desert terrain. The vehicles used GPS, cameras, lasers, and a ton of complicated software to navigate themselves over the course, and in the end, three of them made it. You can watch an awesome hour long NOVA program on the race online here. The next challenge, taking place this year, places the robot cars in an urban environment, where they will have to contend with speed limits, stop signs, and other cars. The vehicles will have to traverse the 60 mile environment in less than 6 hours with no human intervention to win the $3.5 million prize, and they’ll have to obey all California state traffic laws along the way, which is something that most people probably can’t pull off. Danger Room has some excellent coverage of the first few, uh, fatalities. The final will be webcast live on November 3rd, check the DARPA website for details.

[ DARPA Urban Challenge ]

People already love their robots. People already have sex with robots. It’s no stretch of the imagination, therefore, to see humans falling in love with and marrying robots in a few decades, argues David Levy in his Ph.D thesis entitled “Intimate Relationships with Artificial Partners.” Levy explains that humans are already getting quite comfortable interacting with non-human objects, and when you examine the psychological reasons for both love and sex, there’s no real reason why they wouldn’t apply equally to human-human relationships as well as human-robot relationships. His thesis was based on some 450 papers on psychology, sexology, sociology, robotics, materials science, artificial intelligence, gender studies, and computer-human interaction. If the current trends in robot development continue (both in software and, uh, hardware), sooner or later you’re going to end up with a robot that looks and acts arguably just like a human. And even if it’s not just like a human, that may not make a substantial difference when it comes to relationships.

This is basically just what I argued in the article I wrote a few weeks ago about one of the Arse Elektronika talks I attended. It’s funny… A lot of readers were unhappy with the subject matter presented at Arse Elektronika. But when you repackage the concepts that the Arse Elektronika conference was about into an academic format like a Ph.D thesis, it somehow becomes legit. In a lot of ways, the porn industry is way ahead of the game, if for no other reason than they have substantial financial motivation to make sex with robots a mainstream reality. It’s going to happen to society at large at some point in the future, but it’s going to happen FIRST (probably sooner than you might think) in the porn industry, and that’s where we’re going to see the ethical issues initially emerge.

One way or another, it’s gonna be interesting.

[ Daily Mail Article ] VIA [ Crave ]

Natasha from the TED (Technology, Entertainment and Design) Conference has just let us know that a new talk by Hod Lipson on self aware robots is now available for viewing on their website. You can check out the video here (and it is worth watching), but basically Lipson discusses two robot self-modeling projects of his. One of them is the Starfish quadruped robot that’s able to teach itself to walk (we wrote about it a few months ago), and the other project is the Golem Project, which some of you may remember from back in 2000. Read more after the jump. (Read more…)